Memory profiling is used to detect inefficient usage of dynamic memory and memory leaks. Some use cases for the memory profiler are:
- the service is running out of RAM (RSS), which creates problems in production right up to the point of the OOM killer.
- there are suspicions of a memory leak. For example, the RSS usage is constantly growing and does not reach a plateau until the service is restarted.
- there are arguments in favor of the fact that the service will soon experience a shortage of RSS. For example, due to the growth of caches or other internal structures by an order of magnitude.
All the userver based services could use jemalloc profiler that could be is dynamically enabled/disabled via the server::handlers::Jemalloc.
How to profile a running service
- Set the environment variable
MALLOC_CONF
for your service before the service start: MALLOC_CONF="prof:true,prof_active:false,lg_prof_sample:14,prof_prefix:/tmp/jeprof-SERVICE"
- Start your service
- To start sampling call the server::handlers::Jemalloc. If you use the static config from Production configs and best practices then you have to do something like this:
bash
$ curl -X POST localhost:1188/service/jemalloc/prof/enable
OK
$ curl -s -X POST localhost:1188/service/jemalloc/prof/stat | grep ' prof.active:'
prof.active: true
If you see "ОК", then the sampling started.
- After some time make a dump of memory state:
bash
$ curl -X POST localhost:1188/service/jemalloc/prof/dump
OK
- Analyse the dump.
- Do not forget to turn off the profiling:
$ curl -X POST localhost:1188/service/jemalloc/prof/disable
OK
$ curl -s -X POST localhost:1188/service/jemalloc/prof/stat | grep ' prof.active:'
prof.active: false
How to profile at service start
- Set the environment variable
MALLOC_CONF
for your service before the service start: MALLOC_CONF="prof:true,prof_active:true,lg_prof_sample:14,prof_prefix:/tmp/jeprof-SERVICE"
- Start your service
- Ensure that the profiler is active:
bash
$ curl -s -X POST localhost:1188/service/jemalloc/prof/stat | grep ' prof.active:'
prof.active: true
- After some time make a dump of memory state:
bash
$ curl -X POST localhost:1188/service/jemalloc/prof/dump
OK
- Analyse the dump.
- Do not forget to turn off the profiling:
bash
$ curl -X POST localhost:1188/service/jemalloc/prof/disable
OK
$ curl -s -X POST localhost:1188/service/jemalloc/prof/stat | grep ' prof.active:'
prof.active: false
How to analyse the dump
- To decrypt the dump file, you need the binary files of your service and dynamic libraries with which the process was launched during profiling. Therefore, it is recommended to run
jeprof
on the target machine and parse the resulting pdf/text file on your working machine. To install jeprof
, you can do apt install libjemalloc-dev
.
- If you want to get a call graph with notes about the allocated memory, use the command:
bash
$ jeprof --show_bytes --pdf build/services/userver-sample /tmp/jeprof.5503.1.m1.heap> prof.pdf
$ gnome-open prof.pdf # Open the file in PDF viewer
For the command to work, you need to make sure that the apt install graphviz ghostscript
packages are installed.
- If you want to get the top of memory allocation commands, use the command:
bash
$ jeprof --show_bytes --text --cum build/samples/userver-sample /tmp/jeprof.5503.1.m1.heap
Using local file build/services/userver-sample.
Using local file /tmp/jeprof.5503.1.m1.heap.
FAQ